AI reduces time-to-value in advanced process control

One of the key challenges that process engineers face is turning operational data into realized production value. Advanced process control (APC) has become the gold standard for stabilizing complex, multivariable plant environments. However, traditional APC frameworks utilize manual engineering workflows, static modeling and a high degree of reliance on humans.

"The focus of this question is very simple," Jaideep Bhattacharya, global director of offering management at Honeywell, stated during his opening remarks of his presentation at the 2026 Honeywell Users Group in Phoenix. "How do we move faster from data to value by including AI into our APC framework?"

The strategy with Honeywell’s Advanced Process Control suite is not a wholesale replacement of the deterministic algorithms that have safely run plants for decades, noted Bhattacharya. Instead, the goal is to enhance outcomes by including machine learning (ML) and deep learning (DL) into the existing workflow to drive intelligence and accelerate deployment timelines, perhaps even paving a sustainable path toward autonomous operations.

The human bottleneck

APC projects require human intervention across the entire lifecycle. Engineers must manually design the controller architecture, conduct step tests, pre-process historical data and construct model matrices. Once the controller is online, it remains fundamentally static. It relies on operators and process engineers to continuously adjust operational ranges by retuning parameters and establishing new optimization targets based on asset constraints and downstream conditions, said Bhattacharya.

This creates a multivariable problem for human operators who are already managing high-utilization environments. When an operator is overwhelmed, controllers are frequently dropped into manual mode. By embedding supervised and unsupervised machine learning directly into the APC software layer, Honeywell aims to shift these decision-making burdens from the human to the software.

"The idea here is to not completely remove the human, but how do we reduce the dependencies on humans by adapting more AI," Bhattacharya explained. "What it does is not only to reduce the dependency, but it fast-tracks several processes which are needed in the APC workflow."

Production-ready AI capabilities

Honeywell’s software suite already features four core AI-driven tools designed to tackle specific friction points in the APC lifecycle.

1. The Smart APC Supervisor: Operating as an optimization layer positioned above individual controllers, the Smart APC Supervisor orchestrates coordination across multiple distinct APC applications. It bridges the gap between advanced data analytics and real-time control by solving non-linear optimization problems that traditional linear programming cannot effectively manage.

The Smart APC Supervisor relies on a multi-module architecture that includes:

• Clusterization module, which utilizes the hierarchical density-based spatial clustering of applications with noise (HDBSCAN) algorithm, and the system automatically detects shifts in the plant's operation.
• Prediction module, which is built on an extreme gradient boosting (XGBoost) framework; this module predicts process behaviors and adapts online via continuous pre-learning capabilities.
• Optimization engine, in which a differential evolution algorithm calculates optimal targets based on the non-linear XGBoost cost function, passing these optimized targets down to the regulatory layers once validated by the operator.

2. Machine learning and deep learning inferentials: Soft sensors are vital for estimating unmeasured process qualities between delayed laboratory updates. While traditional inferentials rely on linear regression, they often struggle with severe process non-linearities and require extensive expert variable selection. Honeywell's embedded ML inferentials automate data pre-processing, removal of outliers and input selection. The software runs raw historical data through at least six distinct machine learning and deep learning algorithms concurrently.

In testing side-by-side with customer data, deep learning methods consistently yield the lowest error. Engineers do not need data science expertise to deploy these; the software automatically benchmarks the algorithms and recommends the mathematically superior model for online execution.

3. Advanced History ID for accelerated modeling: Model identification is one of the most time-consuming phases of an APC project. Traditional step testing can take weeks and disrupt steady-state operations. Advanced History ID changes this paradigm by mining existing historian data for natural process variations. The tool automatically filters out periods when the plant was down or when valves were saturated or when analyzers were offline. Based on this clean historical data, the system evaluates process interactions to recommend an initial controller design, identifying potential controlled variables, manipulated variables and disturbance variables.

Engineers can input a priori engineering knowledge, such as known mass balance equations or gain ratios, before running the identifier. The tool then outputs an estimated seed model with around 80% accuracy, said Bhattacharya. From there, an automated, closed-loop online stepper fine-tunes the remaining gaps. This approach can compress traditional APC deployment schedules by 30% to 40%, with internal development targets reaching up to a 70% reduction in time-to-value, he noted.

4. The APC Operator Assistant: A frequent point of failure for sophisticated APC applications is operator skepticism. If an operator does not understand why a controller is making a specific multivariable move, the operator might turn it off. The APC Operator Assistant provides explainable AI directly to the console, translating complex matrix math into actionable insights. It answers common questions such as why some variables are moving while others are constrained.

If it hits a hard constraint, the assistant prioritizes a list of five recommended corrective actions based on the calculated impact on the global objective function. For instance, it might suggest widening a specific downstream limit or adjusting a parallel loop.

The operator retains full veto power and can modify a recommendation, such as raising a reflux limit to 240 instead of the software's suggested 250, before hitting accept to download the new boundaries into the active controller, explained Bhattacharya.

Context-aware optimization agents

Looking beyond current availability, Bhattacharya shared details on an emerging, AI-driven Optimization Agent being developed to automatically establish the plant's "golden operating envelope." This agent monitors how operators interact with controllers under changing environmental and economic conditions.

Phase 1 includes context classification by monitoring real-time feed characteristics to detect process shifts. Phase 2 is historical matching, which scans historical records to isolate past periods that match the current process signature. Phase 3 identifies the historical run that yielded the highest objective function value during that specific context to extract a golden envelope. An automated limit is recommended in Phase 4 by extracting the exact variable high/low limits from that peak run and serving them to the operator.

This dynamic constraint management eliminates the need for a process engineer to manually adjust limits when feedstocks change. Initial testing indicates that keeping the controller pinned to this dynamically adjusted golden envelope can improve process yields by up to 3%, said Bhattacharya.

Next steps

Adopting AI simply for the sake of novelty is a misstep to avoid, cautioned Bhattacharya. "The criteria should be linked to some outcome you are looking for,” he said.

For engineering teams looking to capture this accelerated time-to-value, Bhattacharya outlined a roadmap:

1. Audit existing assets: Many plants running recent versions of Honeywell APC software already possess embedded ML inferential and operator assistant capabilities that have simply not been activated. Review current software revisions to see what features can be utilized immediately.
2. Target high-impact variability: Identify unit operations characterized by severe non-linearities, frequent feedstock changes or high operator intervention.
3. Execute narrow pilots: Instead of a full-scale plant overhaul, isolate a single controller bundle. Test the Advanced History ID tool against historical data to benchmark the seed model's accuracy against old manual models.
4. Prioritize the human element: Ensure that, as technology moves closer to autonomy, console operators are trained to interpret AI recommendations. The goal is to keep the operator's interface intuitive, empowering the operator to manage the overall strategy rather than micro-managing individual setpoints.